Suggestion: Make program work natively with YOLO style label (*.txt)

[LeMinhNgan]

Hello,

I find the YOLO style *.txt label has some advantage, for example the labels are still usable if I resize their images. I wish in our program, we can set option to adopt either YOLO *.txt or JSON labels during loading, annotating labels/images natively.

Using converter or exporter is not not a good idea because: to convert YOLO *.txt to JSON we need to read the image size first.

[wkentaro]

Using converter or exporter is not not a good idea because: to convert YOLO *.txt to JSON we need to read the image size first.

Isn't that the good argument why LabelMe's annotation.json is superior format than YOLO annotation.txt?

[LeMinhNgan]

@wkentaro , depends on the section we apply. Let think about Object Detection section. The only benefit of Json is it contains class name inside. Meanwhile the benefit of of YOLO txt is all bounding box coordinates are in ratio 0..1 and independent from the image size. Hence for further process with labels eg. augmentation, splitting image or rename... the YOLO labels can be processed alone without raeding imageshape.
There is another point that i love YOlO txt is: to match the image and its label, we just make sure they have same basename, that is visible in explorer windows; meanwhile for Json style, we have to ensure the image name and pathname inside Json which not visible for quick check

[wkentaro]

Other than the basename argument was not very convincing to me.

basename part can be improved in ui.

[LeMinhNgan]

There is a trick we normally use for labeling large dataset that: we don't label all dataset one time.

1. We only label around 20% of the data and training these labeled data in YOLO
2. Use the trained checkpoint to predict the rest 80% of the dataset --> a label of each image will be generated
3. Attach the predicted labels (*.txt or converted JSON) to the rest 80% dataset
4. Finetune/edit the bounding boxes of those 80% dataset with labeler eg. LabelImg, Lableme.
   --> the difficulty i mentioned is at step 3: we need to convert *.txt labels to JSON label before using them for step 4